Manufacturing AI has to be evaluated as part of a larger system, because improving one model can break other components that have adapted to its errors.
2
Production systems need a defined margin for error, since stopping a manufacturing line can be expensive while missed defects can waste material or create safety risks.
3
Large language models are useful in tightly scoped copilot and administrative tasks, but their changing, probabilistic answers make unsupervised industrial control unsafe.
Summary
The panel describes the practical conditions that make manufacturing AI different from a standard machine learning project. Applications range from battery and semiconductor inspection to battery health, emissions reduction, process optimisation, and safety monitoring. The right target depends on the use case: some systems need very low latency, while others need careful control of false positives, missed defects, or safety failures. The speakers argue that a model cannot be judged alone. It must fit into existing equipment, rule-based checks, human review, fallback models, and production processes. They discuss anomaly detection against an ideal state as a way to catch defects that were not present in the training data. Large language models have useful roles in drafting responses and supporting developers, but the panel is cautious about putting them directly into industrial processes. Repeatability, guardrails, evaluation, and trust matter more than a strong demo.
Manufacturing AI covers process development, inspection, maintenance, and safety
The panel gives examples from several manufacturing settings. Aniket Singh works on battery inspection, including electrode and surface inspection. Jürgen Weichenberger discusses battery health for data-center UPS systems, carbon emissions, digital twins, process optimisation, and electrification. Mohan Mahadevan describes semiconductor process development and volume production. AI helps discover defects that are known, defects that are known to be possible but not fully understood, and defects that have not been seen before.
The right model target depends on the physical process
Jürgen Weichenberger says some computer vision applications have a total decision time of about 50 milliseconds, leaving roughly 10 to 20 milliseconds for inference after image capture and control-system communication. Other applications prioritise accuracy or safety. A false positive can stop a production line, while a missed defect can create large material losses. On a paper machine that is hundreds of metres long, restarting after a stop can take hours. The acceptable tradeoff must be set as part of process design.
Proof of concept must include long validation and a controlled rollout
Aniket Singh says battery manufacturers do not deploy a model simply because it works in an initial test. They spend months checking whether it is robust enough for production because defective batteries can waste material and create risks for vehicles. Before deployment, the team uses secondary inspections rather than letting the model make live decisions. A new model first runs alongside an existing one, and the analytics team compares their performance before production stakeholders approve the change.
The panel rejects the idea that every inspection problem should be solved with AI. Rule-based systems remain useful for dimensions and fast checks. A rule-based detector can identify a possible surface defect, then send only that case to an AI model for confirmation. Other systems use both outputs to make a decision. Jürgen Weichenberger also describes combining physical principles with data-driven models, because a purely optimised solution may leave an engineer with no safety margin.
Anomaly detection can find defects that were never labelled
Jürgen Weichenberger says Schneider Electric chose to model the ideal state rather than enumerate every possible failure pattern. Manufacturing failures can occur in too many forms to provide training examples for all of them, and good production data greatly outnumbers bad data. The system flags deviations from the ideal state, then teams can label new abnormalities and add them to the reference set. Mohan Mahadevan describes this as a system-design problem involving anomaly detection, statistical process boundaries, quality checks, and root-cause analysis.
Pavol Bielik says an AI model is only one component in a larger system. Other components may have learned to compensate for the model's previous mistakes. Replacing it with a model that scores better on average can therefore cause regressions elsewhere. The panel recommends evaluating the complete system, checking corner cases that can stop production, and comparing new behaviour with the existing baseline. Human review remains part of the evaluation, especially for low-confidence images and new defect types.
Industrial language-model use needs narrow scope and strong controls
The panel sees practical value in language models for drafting customer-service replies and supporting developers. These are cases where a person remains in control and can correct a bad answer. Direct industrial use is harder because repeated prompts can produce different instructions, which damages trust for field and service engineers. Mohan Mahadevan says limited production use is possible when the team can control the outcome, while Jürgen Weichenberger argues that guardrails are needed to keep probabilistic systems from changing behaviour in process-control settings.
Model evaluation must test consistency, not only average metrics
Pavol Bielik connects trust to the way a system behaves under small changes. Industrial users need to know that a slight camera shift or a software update will not produce an unexpected decision. The panel says ordinary accuracy and confusion-matrix results are insufficient on their own. Teams also need to inspect examples, test whether existing capabilities remain intact, and build evaluation frameworks around the properties a model should satisfy. Bielik says LatticeFlow is interested in an open-source initiative for this kind of evaluation.
"Continuity in production is more valuable than short periods of peak production and then long prolonged periods of time where you are not running at all because you break something."Jürgen Weichenberger25:44
Who should watch
You are taking a computer vision or inspection model from a factory proof of concept toward a live production line and need to plan validation, fallback behaviour, and approval steps.
Your team is deciding where AI should replace an existing rule-based check and where a hybrid system is safer or faster.
You are considering language models for industrial operations and need a sober account of repeatability, human review, guardrails, and trust.
